A Statistical Learning Approach to Human Behavior and Status Inference for Unmanned Vehicle Operation

Abstract

therefore, the sparseness of the simulation log is compensated and the latent factors that explain the decisions and behaviors of experienced operators are revealed. From comparison experiments, the proposed method outperformed the other methods. It is expected that the proposed method will contribute to the development of autonomous vehicles that perform as well as human operators. The contribution and utility of this dissertation are summarized into three points. Most importantly, this dissertation covers timely issues related to UVs, which are essential both to reducing the accident rate and to enhancing performance of UVs. Second, previous research regarding inferring the status and behavior of UV operators and related domains were thoroughly summarized to provide a solid background to the problems. Finally, with some adjustments, the proposed methods are able to be extended to related domains that utilize heterogeneous data.

Recently, technologies for unmanned vehicles (UVs) are getting increased attention because of the increasing popularity of UVs. A UV is a vehicle without an on-board operator, and, therefore, the utilization of UVs leads to cost reduction for operations and better safety for operators. UVs are categorized into remotely operated vehicles and autonomous vehicles according to whether a human operator is required. For the successful operation of remotely operated vehicles, it is important to detect the inattention status of an operator because the occurrence of inattention is known to be the main cause of accidents for UVs. Accurate reproduction of a behavior and extraction of an optimal behavior are the main problems for autonomous vehicles. This dissertation aims to address the problems in UVs by adopting statistical learning methods. With the advances in sensor and data storage technologies, a large number of datasets generated by UVs, such as EEG signals of the operator and simulation log for combat, has become available. However, these datasets contain heterogeneous data composed of multi-type attributes, making the application of statistical learning challenging. The methods proposed in this dissertation are as follows. First, a method for inferring the mental status of UV operators from EEG signals is introduced. A semi-supervised learning method that utilizes an attribute-weight learning algorithm is proposed, where the attention duration at the beginning of an operation and the dierent levels of correlations between attributes and labels are exploited. As a result, occurrences of operator inattention during maneuvering of UVs were successfully detected when applying the proposed method to experiments using a real-world dataset. Second, a method for inferring the behavior of UV operators from a simulation log is presented. A hierarchical support vector machine with a hybrid sequence kernel is presented to address the heterogeneity of the considered dataset in terms of value type and sequence dependency, and to enhance performance by incorporating a hierarchical structure of behaviors of UV operators. Finally, a method for inferring the optimal behavior for UV operators from a simulation log is developed. The proposed method is based on a matrix factorization algorithm